Figures (10)  Tables (5)

    • Figure 1. 

      Results of data processing.

    • Figure 2. 

      Structure of YOLOv8.

    • Figure 3. 

      Structure of coordinate attention.

    • Figure 4. 

      Structure of the improved YOLOv8.

    • Figure 5. 

      Traffic sign categories.

    • Figure 6. 

      Number of traffic signs for each category.

    • Figure 7. 

      Structure of Jetson Nano.

    • Figure 8. 

      Detection results of YOLOv8-CE under different weather test sets.

    • Figure 9. 

      Hardware system of Jetson Nano.

    • Figure 10. 

      Field test on Jetson Nano in Harbin, China.

    • Parameter Technical specifications
      AI performance 472 GFLOPs
      GPU 128-core Maxwell
      Memory 4 GB 64-bit LPDDR4 25.6 GB/s
      CPU Quad-core ARM A57 @ 1.43 GHz
      CPU max frequency 1.43 GHz
      Storage microSD (not included)
      Connectivity Gigabit Ethernet, M.2 Key E
      USB 4x USB 3.0, USB 2.0 Micro-B
      Power 5 W - 1 0W

      Table 1. 

      Hardware parameters of Jetson Nano.

    • Model Precision Recall mAP@0.5
      YOLOv8n 88.7% 73.6% 83.3%
      CA 89.6% 75.4% 84.9%
      EIoU 87.9% 75.1% 84.5%
      CA + EIoU 90.2% 78.1% 86.1%

      Table 2. 

      Results of the ablation study.

    • Model Precision Recall mAP@50 mAP@50-95 Inference time (ms)
      YOLOv8-CE(ours) 90.2 78.1 86.1 57.2 96
      YOLOv8n 88.7 73.6 83.3 53.7 92
      YOLOv8-ghost 89.2 72.5 82.1 52.1 243
      YOLOv8-ghostv2 89.1 71.6 82.2 52.5 230
      YOLOv8-shufflenetv2 81.1 53.6 61.8 61.8 110
      YOLOv8-mobilenetv3 73.4 54.6 61.4 35.7 65

      Table 3. 

      Comparison of traffic sign detection models.

    • Model Orinal Cloud Foggy Night Rain Snow Sunny
      YOLOv8-CE(ours) 86.1 92.5 81.6 76.5 43.1 86.5 94.8
      YOLOv8n 83.3 89.5 68.3 73.8 32.7 77.8 93.2
      YOLOv8-ghost 82.1 88.8 77.9 75.5 29.8 82.5 91.3
      YOLOv8-ghostv2 81.9 89.2 63.9 73.0 39.9 70.0 91.3
      YOLOv8-shufflenetv2 61.4 74.9 53.7 40.6 10.8 45.7 80.0
      YOLOv8-mobilenetv3 61.8 75.9 55.2 40.6 16.9 58.6 77.6

      Table 4. 

      Comparison of traffic sign detection models under different weather test sets.

    • Model
      		FLOPs
      (G)      		 Weight size
      (MB)      		 Inference time (ms)
      Jetson Nano Raspberry pi 4B
      YOLOv8-CE(ours) 8.1 5.99 96 690
      YOLOv8n 8.1 5.96 92 678
      YOLOv8-ghost 6.8 5.97 243 810
      YOLOv8-ghostv2 6.8 5.17 230 791
      YOLOv8-shufflenetv2 5.0 3.48 110 580
      YOLOv8-mobilenetv3 2.8 2.52 65 313

      Table 5. 

      Experiments on different devices.